particle flow and ilc detector designthomson/talks/alcpg_fermilab.pdf · alcpg seminar, fermilab...

ALCPG Seminar, Fermilab 8/12/2006 Mark Thomson 1

Particle Flow and Calorimetry at the ILC

Mark ThomsonUniversity of Cambridge

This Talk:ILC Physics ILC CalorimetryIntroduction to PFACalorimetry in the ILC Detector ConceptsPFA and Detector DesignPandoraPFACurrent Performance Detector Optimisation StudiesCurrent LimitationsOutlookConclusions


ILC Physics ↔ CalorimetryPrecision Studies/Measurements

Higgs sectorSUSY particle spectrumSM particles (e.g. W-boson, top)and much more...

ILC PHYSICS:

•ZHH

Require High Luminosity Detector optimized for precision measurements

in difficult multi-jet environment

σ(e+e- ZHH) = 0.3 fbe.g.Small cross-sections

High Multiplicity final statesoften 6/8 jets

Physics characterised by:

Some preliminary

Compare with LEPe+e- W+W-e+e- Z and dominate

backgrounds not too problematicKinematic fits used for mass reco.

good jet energy resolution not vital

Physics performance depends critically on the detectorperformance (not true at LEP)Places stringent requirements on the ILC detector

At the ILC:Backgrounds dominate ‘interesting’ physicsKinematic fitting much less useful: Beamsstrahlung +final states with > 1 neutrino


Best at LEP (ALEPH):σE/E = 0.6(1+|cosθJet|)/√E(GeV)

ILC GOAL:σE/E = 0.3/√E(GeV)

Jet energy resolution:

σE/E = 0.6/√E σE/E = 0.3/√E

Reconstruction of twodi-jet masses allows discrimination of WWand ZZ final states

If the Higgs mechanism is not responsible for EWSB then QGC processes important

e+e- ννZZ ννqqqqe+e- ννWW ννqqqq ,

THIS ISN’T EASY !

Often-quoted Example:Jet energy resolution directly impacts physics sensitivity

Calorimetry at the ILC

NOTE: this is fastsimulation not full MC

EQUALLY applicable to any final states where want to separateW qq and Z qq !



The Particle Flow ParadigmMuch ILC physics depends on reconstructing jet-jet invariant massesOften kinematic fits won’t help – Unobserved ν, Beamsstrahlung, ISRAim for jet energy resolution ~ ΓZ for “typical” jetsIf we assume σE/E = α/√E(GeV)

Di-jet mass resolution is approx. σm/m = α/√Ejj(GeV)For typical ILC jet pair energies (200 GeV)

σE/E ~ 0.3/√E(GeV)Jet energy resolution is the key to calorimetry at the ILCWidely believed that PARTICLE FLOW is the best way to achieve this

The Particle Flow Analysis (PFA): • Reconstruct momenta of individual particles avoiding double

countingCharged particles in tracking

chambersPhotons in the ECALNeutral hadrons in the HCAL

(and possibly ECAL)

Need to separate energy deposits from different particlesNot calorimetry in the traditional sense


granularity more important than energy resolution

Component Detector Frac. of jet energy

Particle Resolution

Jet Energy Resolution

Charged Particles (X±) Tracker 0.6 10-4 EX neg.

Photons (γ) ECAL 0.3 0.11√Eγ 0.06√Ejet

Neutral Hadrons (h0) HCAL 0.1 0.4√Eh 0.13√Ejet

TESLA TDR achieved resolution for Z uds at rest of ~0.30√Ejet

Energy resolution gives 0.14√Ejet (dominated by HCAL)

Calorimetric performance not the limitation !

In addition, have contributions to jet energy resolution due to “confusion”, i.e. assigning energy deposits to wrong reconstructed particles. This leads to double-counting or incorrectly merging neutrals in to charged showers

σjet2 = σx±2 + σγ

2 + σh02 + σconfusion

2 + σthreshold2 +…

Single particle resolutions not the dominant contribution to jet energy res.

PFA : Basic issuesWhat are the main issues for PFA ? Separate energy deposits + avoid double counting

e.g.Need to separate “tracks” (charged hadrons) from photons

γ


+software

γ

granularity

Need to separate neutral hadrons from charged hadrons

Granularity helpsBut less clear…

Isolated neutral hadron orfragment from shower ?


PFA : “Figure of Merit”For good jet energy resolution need to separate energy deposits from different particles

Large detector – spatially separate particlesHigh B-field – separate charged/neutralsHigh granularity ECAL/HCAL – resolve particles

Physics argues for : large + high granularity + BCost considerations: small + lower granularity + B

R

d=0.15BR2/pt

Often quoted* “figure-of-merit”: BR2

σ

Separation of charge/neutrals

Calorimeter granularity/RMoliere

HIGHCOST

Need realistic algorithms to determine what drives PFA performance….

*But almost certainly wrong (see later)

The ILC Calorimeter Concepts

LDC : Large Detector Concept(spawn of TESLA TDR)

SiD : Silicon Detector

GLD : Global Large Detector

ILC Detector Design work centred around 4 detector “concepts”Each will contribute to an ILC detector conceptual design report by end of ~2006Ultimately may form basis for TDRs3 of these concepts “optimised”for PFA Calorimetry SiD, LDC, GLD

ILC Detector Concepts:

4



Main Differences:

Tracker ECAL

SiD

LDC

GLD

B = 5TB = 4T

B = 3T

SIZE + B-Field

Central Tracker and ECALSiD LDC GLD

TrackerECAL

Silicon TPC TPCSiW SiW Pb/Scint

SiD + LDC + GLD all designed for PFA Calorimetry !also “4th” concept designed for more “traditional” approach to calorimetry !


LDC/SiD CalorimetryECAL and HCAL inside coil

HCAL

ECAL

ECAL: silicon-tungsten (SiW) calorimeter:• Tungsten : X0 /λhad = 1/25, RMoliere ~ 9mm

(gaps between Tungsten increase effective RMoliere)• Lateral segmentation: ~1cm2 matched to RMoliere

• Longitudinal segmentation: 30 layers (24 X0, 0.9λhad)

• Typical resolution: σE/E = 0.15/√E(GeV)

Very high longitudinal and transverse segmentation

Two Main Options:Tile HCAL (Analogue readout)Steel/Scintillator sandwich Lower lateral segmentation

~ 3x3 cm2 (motivated by cost)Digital HCALHigh lateral segmentation

~ 1x1 cm2

digital readout (granularity)RPCs, wire chambers, GEMS…

Again Highly Segmented – for Particle Flow• Longitudinal: ~40 samples• 4 – 5 λ (limited by cost - coil radius)• Would like fine (1 cm2 ?) lateral segmentation• For 10000 m2 of 1 cm2 HCAL = 108 channels – cost !

Hadron Calorimeter

The Digital HCAL Paradigm

p

Only sample small fraction of the total energy deposition

• Sampling Calorimeter:

• Energy depositions in activeregion follow highly asymmetric Landau distribution OPEN QUESTION


GLD CalorimetryECAL and HCAL inside coil Tungsten

Scintillator

4mm 2mm

W-Scintillator ECALsampling calo.

Initial GLD ECAL concept:

Strips : 1cm x 20cm x 2mm

Achieve effective ~1cm x 1cm segmentation using strip/tilearrangement

Tiles : 4cm x 4cm x 2mm

Big question of pattern recognitionin dense environment

Pb-Scintillator HCALsampling calo.

SiD/LDC/GLD : Basic design = sampling calorimeter



Calorimeter ReconstructionHigh granularity calorimeters –very different to previous detectors(except LEP lumi. calorimeters)

“Tracking calorimeter” – requiresa new approach to ECAL/HCALreconstruction

+PARTICLE FLOW

ILC calorimetric performance = HARDWARE + SOFTWARE

Performance will depend on the software algorithm

Nightmare from point of view of detector optimisation

a priori not clear what aspects of hadronic showersare important (i.e. need to be well simulated)


PFA and ILC detector design ?

VTX : design driven by heavy flavour tagging,machine backgrounds, technology

PFA plays a special role in design of an ILC Detector

ECAL/HCAL : single particle σE not the main factorjet energy resolution ! Impact on particle flow drives

calorimeter design + detector size, B field, …

Tracker : design driven by σp, track separation

PFA is a (the?) major $$$ driver for the ILC Detectors

BUT: Don’t really know what makes a good detector for PFA(plenty of personal biases – but little hard evidence)

How to optimise/compare ILC detector design(s) ?Need to choose the key “benchmark” processes (EASY)


For example:The rest is VERY DIFFICULT !

e.g. tt event in LDC e.g. tt event in SiD

Wish to compare performance of say LDC and SiD detector concepts

However performance = DETECTOR + SOFTWARENon-trivial to separate the two effectsNEED REALISTIC SIMULATION + REALISTIC RECONSTRUCTION !- can’t use fast simulation etc.


PandoraPFA*Need sophisticated reconstruction before it ispossible to start full detector design studies

So where are we now ?

Significant effort (~6 groups developing PFA reconstruction worldwide)

For this talk concentrate on: PandoraPFAThis is still work-in-Progress – but does a pretty good job+ beginning to get a better feel for what really matters…. Will give a fairly detailed description of the algorithm andhighlight the short-comings

Then discuss some first detector optimisation studies

*Born in the USA : Snowmass 2005


What software is needed?

LCIO DATA

G4 Simulation

Digitisation

Tracking

VertexingFlavour Tag

Clustering Particle Flow

Framework

Physics Analysis

What exists now?

Mokka

“Simple” Digitisers

TrackCheater

LEP TPC Tracking LDC Tracking

ZVTOPANN

PandoraPFA

Wolf

MARLIN

European Software Framework


PandoraPFA OverviewECAL/HCAL reconstruction and PFA performed in a single algorithmKeep things fairly generic algorithm

applicable to multiple detector conceptsUse tracking information to help ECAL/HCAL clustering

This is a fairly sophisticated algorithm : ~8000 lines of code Will discuss this in some detail – illustrates practical issuesfor PFA reconstruction

Six Main Stages (soon to be seven):i. Preparation ii. Loose clustering in ECAL and HCALiii. Topological linking of clearly associated clustersiv. Courser grouping of clustersv. Iterative reclusteringvi. Formation of final Particle Flow Objects

(reconstructed particles)


Preparation I: Extended HitsCreate internal ExtendedCaloHits from CaloHitsExtendedCaloHits contain extra info:

pointer to original hit pseudoLayer (see below)measure of isolation for other hitsis it MIP like actual layer (decoded from CellID)Pixel Size (from GEAR) – hits are now self describing

Hit Hit

YES NO

Arrange hits into PSEUDOLAYERS i.e. order hits in increasing depth within calorimeterPseudoLayers follow detector geometry

Small physical layer no.

High pseudolayer no.i.e. deep in calo.


Preparation II: Isolation

Divide hits into isolatedand non-isolatedOnly cluster non-isolatedhits“Cleaner”/Faster clusteringSignificant effect for scintillator HCAL

Removal of isolated hitsdegrades HCAL resolutione.g. LDC50 %/√E/GeV 60 %/√E/GeV

Preparation III: Tracking

Use MARLIN TrackCheaterTracks formed from MC Hits in TPC/FTD/VTXSimple Helix Fit ⇒ track paramsCuts (primary tracks):

|d0| < 5 mm|z0| < 5 mm>4 non-Si hits

+ V0 and Kink finding:

Improves PFA performanceby ~2 %

Track resolution better thancluster

Will soon move to fully reconstructed tracks (LDCTracking)


ii) ECAL/HCAL Clustering Start at inner layers and work outwardTracks can be used to “seed” clustersAssociate hits with existing ClustersIf no association made form new ClusterSimple cone based algorithm

Simple cone algorithmbased on current direction+ additional N pixels

Cones based on either:initial PC direction orcurrent PC direction

0 1 2 3 4 5 6

Unmatched hits seeds new cluster

Initial clusterdirection

Parameters:cone angleadditional pixels



iii) Topological Cluster AssociationBy design, clustering errs on side of caution

i.e. clusters tend to be splitPhilosophy: easier to put things together than split them upClusters are then associated together in two stages:

• 1) Tight cluster association – clear topologies• 2) Loose cluster association – fix what’s been missedPhoton ID

Photon ID plays important role Simple “cut-based” photon ID applied to all clustersClusters tagged as photons are immune from associationprocedure – just left alone

γ γγ

Won’t mergeWon’t merge Could get merged


• Join clusters which are clearly associated making use of high granularity + tracking capability: very few mistakes

Clusters associated using a number of topological rules Clear Associations:

Less clear associations:

Proximitye.g.7 GeV cluster

4 GeV track

6 GeV cluster

Use E/p consistency to veto clear mistakes


Topological association : track merging

LOOPERSTight cut on extrapolation ofdistance of closest approachof fits to ends of tracks

SPLIT TRACKSgap

Tight cut on extrapolation ofdistance of closest approachof fits to end of inner tracksand start of outer track


Topological Association II : BackscattersForward propagation clustering algorithm has a major drawback:back scattered particles form separate clusters

Project track-like clusters forwardand check distance to shower centroidsin subsequent N layers

Also look for track-like segments at startof cluster and try to match to end of another cluster


Topological association III : MIP segmentsLook at clusters which are consistent with having tracks segmentsand project backwards/forward (defined using local straight-line fitsto hits tagged as MIP-like)

Apply tight matching criteria on basis of projected track[NB: + track quality i.e. chi2]


iv) Cluster Association Part II• Have made very clear cluster associations• Now try “cruder” association strategies• BUT first associate tracks to clusters (temporary association)• Use track/cluster energies to “veto” associations, e.g.

5 GeV track

6 GeV cluster

7 GeV cluster

This cluster association would beforbidden if |E1 + E2 – p| > 3 σE

Provides some protection against “dumb” mistakes


Distance betweenhits : limited to firstpseudo-layers of cluster

Proximity

Shower Cone

Associated if fraction ofhits in cone > some value

Shower start identified

Apply looser cuts if have low E clusterassociated to high E track+Track-Driven Shower Cone


v) Iterative Reclustering Upto this point, in most cases performance is good –but some difficult cases…

30 GeV π+

20 GeV n

At some point hit the limit of “pure” particle flowjust can’t resolve neutral hadron in hadronic shower

e.g. if have 30 GeV track pointing to 20 GeV clusterSOMETHING IS WRONG

The ONLY(?) way to addressthis is “statistically”


18 GeV

If track momentum and cluster energy inconsistent : RECLUSTERe.g.

30 GeV 12 GeV

10 GeV Track

Change clustering parameters until cluster splits and get sensible track-cluster match

NOTE: NOT FULL PFA as clustering driven by track momentum

This is very important for higher energy jets


Iterative Reclustering StrategiesCluster splitting

Reapply entire clustering algorithm to hits in “dubious” cluster. Iterativelyreduce cone angle until cluster splits to give acceptable energy match to track

30 GeV12 GeV

18 GeV

10 GeV TrackCould plug in alternative clustering

Cluster merging with splitting

30 GeV Track

Look for clusters to add to a track toget sensible energy association. If necessary iteratively split up clusters to get good match.

38 GeV 18 GeV

12 GeV 32 GeV

Track association ambiguitiesIn dense environment may have multiple tracks matched to same cluster. Apply above techniques to get ok energy match.

“Nuclear Option”If none of above works – kill track and rely on clusters alone


Current Performance

σE/E = α√(E/GeV)EJET

180 GeV 0.57

|cosθ|<0.8

45 GeV 0.30

100 GeV 0.37

250 GeV 0.75

Find smallest region containing90 % of eventsDetermine rms in this region

Figures of Merit:

rms90 σ75

Fit sum of two Gaussians with same mean. The narrower one is constrained to contain 75% of events Quote σ of narrow Gaussian

It is found that rms90 ≈ σ75

For jet energies < 100 GeVperformance is probably good enough for physics studies


The current performance of the algorithm is well described by theEMPIRICAL expression:

Nothing deep herejust current state of play

Angular Dependence

• Jet energy resolution depends onpolar angle

• Degradation in endcap : nuclearinteractions in TPC endplate have someimpact + longer track extrapolation

• For high energy jets performance in barrel region worse at low values of |cosθ| - leakage (see later)

+ HCAL ring not currently simulated in Mokka


Recent Detector Optimisation Studies From point of view of detector design – what do we want to know ?

Optimise performance vs. cost

Main questions (the major cost drivers):• Size : performance vs. radius• Granularity (longitudinal/transverse): ECAL and HCAL• B-field : performance vs. B

To answer them use MC simulation + PFA algorithm

!• Need a good simulation of hadronic showers !!!• Need realistic PFA algorithm

(want/need results from multiple algorithms)

This is important – significant impact on overall design ofxxx M$ detector !

Interpretation of results needs care – observing effects of detector+ imperfect software


PFA-related Detector Design issues What aspects of the detector might impact PFA performance?

Main questions identified at Snowmass (in some order of priority):1) B-field : Does B help jet energy resolution2) Size : ECAL inner radius/TPC outer radius3) TPC length/Aspect ratio4) Tracking efficiency – forward region5) How much HCAL – how many interactions lengths 4, 5, 6…6) Longitudinal segmentation – pattern recognition vs sampling

frequency for calorimetric performance7) Transverse segmentation ECAL/HCAL

ECAL : does high/very high granularity help ? 8) Compactness/gap size9) Impact of dead material10) How important are conversions, V0s and kinks 11) HCAL absorber : Steel vs. W, Pb, U…12) Circular vs. Octagonal TPC (are the gaps important)13) HCAL outside coil – probably makes no sense but worth

demonstrating this (or otherwise)14) TPC endplate thickness and distance to ECAL15) Material in VTX – how does this impact PFA


e.g. B-Field at 91.2 GeVLDC00 Detector (≈ TESLA TDR) – same event different B

B = 2T B = 4T B = 6T

σE/E = α√(E/GeV)B-Field

All angles34.1±0.3%33.4±0.3 %

|cosθ|<0.72 Tesla 30.8±0.4 %4 Tesla 29.2±0.4 %6 Tesla 34.4±0.3 % 29.7±0.4 %

Only weak B-fielddependence

The reason being thatconfusion is not dominatingthe resolution for these lowenergy jets


Radius vs FieldNow for a more serious study…Map out the dependence on B and outer radius of the TPCUse LDC00 detector model with:

rtpc = 1380-2280 mmB = 3-5 Tesla

rTPC = 1380 mm rTPC = 1580 mm rTPC = 1690 mm rTPC = 1890 mm

• Look at jet energy resolution for Z uds events at• √s = 200 GeV• √s = 360 GeV


SiD-likeLDC

GLD-like

Results consistent with:

LDC00 180 GeV jets100 GeV jets

Empirical

As expected large radius/ large field does bestBut not as strong an effect as might have been expectedHow much due to “intrinsic detector resolution” and how muchdue to software deficiencies ?


HCAL Depth and Transverse segmentationInvestigated HCAL Depth (interaction lengths)• Generated Z uds events with a large HCAL (63 layers)

• approx 7 λI• In PandoraPFA introduced a configuration variable

to truncate the HCAL to arbitrary depth• Takes account of hexadecagonal geometry

• For 100 GeV Jets no advantage in going to larger HCAL !

• For 180 GeV Jets HCAL leakage degrades PFAperformance

4.3 λI 5.3 λI

NOTE: no attempt to account for leakage – i.e. using muon hits - this is a worse case


Analogue scintillator tile HCAL : change tile size 1x1 10x10 mm2

1x1 3x3 5x5 10x10

σEvis/E = α√(E/GeV)Detector Model

Z @91 GeV tt@500 GeV

LDC00Sc 1cm x 1cm 31.4 ± 0.3 % 42 ± 1 %LDC00Sc 3cm x 3cm 30.6 ± 0.3 % 45 ± 1 %

31.3 ± 0.3 % 48 ± 1 %56 ± 1 %33.7 ± 0.3 %

LDC00Sc 5cm x 5cmLDC00Sc 10cm x 10cm

10x10 too coarse (can be seen clearly from display)Finer granularity helps(?) somewhat at higher energies maybe better tracking of overlapping showers?

Visible energyresolution


ECAL Transverse Granularity

• Detector model:LDC00Sc (~Tesla TDR)

• B = 4 Tesla• 30x30mm2 HCAL

• Use Mokka to generate Z uds events @ 200 GeV with differentECAL segmentation: 5x5, 10x10, 20x20 [mm2]

• For 100 GeV jets, not a big gain going from 10x10 5x5mm2

With PandoraPFA

[ for these jet energies the contributions from confusion inside the ECAL is relatively small – need ]

• 20x20 segmentation looks too coarse

For a small detector: finer granularity does help


Caveat EmptorThese studies are interesting but not clear how seriously theyshould be taken

how much is due to the detectorhow much due to imperfect algorithm

Need results from other algorithms

So what are the current deficiencies of the algorithm…?


Current Limitations Matrix of Confusion: look MC at fractions of the total energy generated in the different particle types (h±, γ, h0) and compare to the reconstructedh±, γ, h0 fractionsFor perfect reconstruction this would be diagonale.g. 100 GeV uds jets (all polar angles)

h± γ h0

h±

γ

h0

58.2% 0.4% 3.5%

0.6% 19.5% 2.7%

1.8% 0.9% 12.4%

Generated as

Rec

onst

ruct

ed a

s

Confusion between photons and neutral hadrons not critical

Fragment clusters Which should have been associated to a track

Merged clusters

Fragment clusters are the biggest single problem


Fragments A few example events :

Green = correctly identified neutral clustersRed = reconstructed neutral clusters from charged hadron


Work in progress to explicitly identify these fragments as the laststage in the reconstruction based on:

cluster shape, cluster direction, …Expect a not insignificant improvement in PFA performance

Also issues with tracking which will improve with new full trackreconstruction (better track extrapolation will help)


OutlookPandora Code was “released” in the Summer:

http://www.hep.phy.cam.ac.uk/~thomson/pandoraPFAStill some work to do…Also working to make code compatible with SLIC and Jupiterevents – clustering works well, some issues with tracking PandoraPFA is not perfect, but does a reasonable job for jetenergies < 100 GeVCan start to use it for full simulation physics studies

Z

Z

e.g. “classic plot”

Full simulation/PFA reconstruction


ConclusionsGreat deal of effort (worldwide) in the design of the ILC detectorsCentred around 4 “detector concept” groups: GLD, LDC, SiD + 4th

Widely believed that calorimetry and, in particular, jet energy resolution drives detector designAlso believed that it is likely that PFA is the key to achieving ILC goal

Calorimetry at the ILC = HARDWARE + SOFTWARE (new paradigm)It is difficult to disentangle detector/algorithm…. Can only address question with “realistic algorithms”

i.e. serious reconstruction 10+ years before ILC turn-onWith PandoraPFA algorithm already getting to close to

ILC goal (for Z uds events)More importantly, getting close to being able to address real issues:

What is optimal detector size/B-field, etc.

THIS IS HARD – BUT VERY IMPORTANT !

GLD, LDC, SiD calorimetry “designed” for PFANeed to demonstrate this actually makes sense !

not yet completely proven…!Need to study in context of physics sensitivity

FINAL COMMENT:


End

particle flow and ilc detector designthomson/talks/alcpg_fermilab.pdf · alcpg seminar, fermilab...

Documents